Optimising feature weights

mbollmann / perceptron

Collection of various perceptron algorithms, mainly for NLP purposes

MIT License

0 stars 0 forks source link

Open mbollmann opened 9 years ago

mbollmann commented 9 years ago

do not grow weights during prediction (unknown features will contribute nothing anyway)
prune dimensions that are zero (or very close to it) from the learned weights
alternatively: prune features that only occurred less than x times (probably requires keeping count)

mbollmann commented 8 years ago

The first two suggestions have been implemented in commit a6b45db.

Still leaving this open since the third suggestion also seems worth exploring.